ludwig-ai--ludwig/examples/llm_config_generation

LLM-Driven Config Generation

Note: This feature requires PR #4092 to be merged into Ludwig, or pip install ludwig>=0.14.

What is this?

Ludwig's config generation feature lets you describe your machine learning task in plain English and receive a fully validated Ludwig configuration file in return. An LLM (Claude or GPT-4) interprets your description, maps column names to Ludwig feature types, selects an appropriate model architecture, and emits a config dict that passes Ludwig's Pydantic schema validation before it ever reaches your code.

This is particularly useful for:

• New users who are unfamiliar with Ludwig's YAML schema and want a working starting point.
• Rapid prototyping — describe the task, inspect the generated config, tweak if needed, and run.
• Multi-task problems — describing simultaneous outputs (e.g. classify + regress) is often easier in prose than in YAML.

Prerequisites

You need an API key for at least one of the supported backends:

Backend	Environment variable
Anthropic (Claude)	ANTHROPIC_API_KEY
OpenAI (GPT)	OPENAI_API_KEY

The library reads the key automatically from the environment. You can also pass api_key= explicitly.

Install the required packages:

pip install "ludwig>=0.14" anthropic   # for Claude
# or
pip install "ludwig>=0.14" openai      # for GPT

Quick start

import os
import yaml
from ludwig.config_generation import generate_config  # requires PR #4092 / ludwig>=0.14

config = generate_config(
    "I have customer data with age, income, and purchase history. "
    "I want to predict churn (binary) and lifetime value (number).",
    model="claude-sonnet-4-20250514",
    # api_key is read from ANTHROPIC_API_KEY by default
    validate=True,
)

print(yaml.dump(config, default_flow_style=False))

You can also use an OpenAI model by passing its name:

config = generate_config(
    "Predict apartment rent price from sqft, bedrooms, and neighborhood.",
    model="gpt-4o",
    validate=True,
)

The backend is chosen automatically based on whether the model name starts with "claude" or "gpt".

Files

File	Description
README.md	This file
llm_config_generation.ipynb	Interactive walkthrough notebook
generate_and_train.py	Standalone CLI script — describe a task, confirm, train

Running the standalone script

# Use the default task description
python generate_and_train.py

# Or pass your own description
python generate_and_train.py "predict house price from bedrooms, sqft, and location"

# Use a specific model
python generate_and_train.py --model gpt-4o "classify email sentiment as positive, neutral, or negative"

Tips for writing good task descriptions

• Name your columns — "age, income, and purchase_count" is more actionable than "some user features".
• State the target and its type — "predict churn (binary)" or "predict revenue (continuous number)".
• Mention modalities — "text product description and tabular price, category" helps Ludwig pick the right encoder.
• Include rough dataset size — "~50 k rows" lets the LLM suggest appropriate model complexity.
• Describe multi-output tasks explicitly — "simultaneously predict price (regression) and category (classification)".